Self Standard
thealialexander/self ...

Version 0.5 proposal

You should be able to ask your AI about itself.

The Self standard is a schema for model identity, orchestration and harness disclosure, and user preferences. It allows the AI to become self referential.

Explore schema Customize
self/
SKILL.md SELF.md orch.md model.md user.md harness.md skills/ scripts/ security.md chrono/ metadata.json manifest.json

Built with You in mind

Self is not another giant prompt file.

It is a small canonical entrypoint plus scoped Markdown slots. Each file starts with Skills-style frontmatter, then keeps descriptions, procedures, tables, and examples in the body for triggered loading.

Map first

Agents start with self/SELF.md, then decide whether deeper files are relevant.

Markdown canonical

Humans can read or import the folder without a parser. The manifest helps tools validate it.

Shaped like Skills

Metadata is cheap and always indexable. Instructions and resources load only after a trigger.

Privacy built-in

user.md and chrono/ are consent-gated and excluded from shared context unless the user asks.

Harness aware

The assistant can understand its app surface, scripts, tools, permissions, and common failure modes.

Schema explorer

Open only the slot the task needs.

Choose a slot to see its Skills-style frontmatter, loading rule, and example content. The explorer mirrors the starter folder embedded in this page.

Public

Self

summary

self/SELF.md

Primitive

Progressive disclosure is normative.

The standard treats context as a budget and a trust boundary. A conforming agent can explain why it opened a file, and why it left the rest alone.

1

Read map

self/SELF.md lists slots, privacy, summaries, and open conditions.

2

Select slot

The task, user request, or harness error chooses the next file.

3

Open detail

Full Markdown is loaded only when its frontmatter and slot catalog permit it.

4

Report use

The agent can state which Self files influenced its answer.

Local validator

Embedded starter checks

These checks run in the browser against the examples embedded in app.js. No network or server is used.

For Agent Skills conformance, run the official validator locally: uvx --from skills-ref agentskills validate examples/self.

    Example questions

    Route the question before loading the memory.

    These examples show how Self answers identity, environment, continuity, belief, metadata, and app-control questions without opening every file.

    Install Self

    Download the folder, then give it to your agent.

    Start with the full starter zip so self/SELF.md, private placeholder slots, security policy, native Skill wrappers, and adapter templates are all present together.

    1

    Download

    Install the complete starter package first. It includes the self/ folder and harness bridge examples.

    2

    Connect

    After unzipping the package, paste this into your agent or ask it to follow self/skills/first-use.md.

    Codex instruction
    3

    Advanced bridge file

    The path below is inside the downloaded starter zip. In this repo, the source files live under examples/codex/ and examples/claude/.

    adapters/codex/AGENTS.md
    self/skills/first-use.md

    Developers: Extend Self to your App

    The JSON is the machine-readable table of contents.

    Markdown stays canonical for humans and works great for LLM Skills. But, manifest.json lets tools discover, validate, index, package, and selectively load Self without opening every file. This allows Self to be built into other applications that use AI.

    01

    Discovery

    Find the entrypoint, supported version, profile, metadata format, and available slots.

    02

    Validation

    Check that declared paths exist, privacy flags are present, and Markdown files have frontmatter.

    03

    Selective loading

    Index cheap metadata first, then open one file or resource only when a trigger matches.

    04

    Consent-gated context

    After Self is installed, keep private slot contents out of model context unless the user asks for personalization, beliefs, or continuity.

    Developers

    Integrate Self as a package, inside your service.

    Apps can treat manifest.json as a machine-readable index, route requests to the smallest relevant slot, and serve per-user Self state without prompt dumping.

    Download package

    Start from the same complete starter package, then replace placeholder files with app-specific and user-specific state.

    Fill app-owned slots

    Populate model.md, harness.md, and orch.md from your app: model creator/provider facts, runtime surfaces, permissions, memory assembly, and fallback behavior.

    Route by manifest

    Load manifest.json in a service layer, index names and descriptions, then retrieve only SELF.md plus the selected slot for each agent request.

    Serve each user

    Store user.md and chrono/ per user with explicit read/write rules. Load them only for personalization, beliefs, continuity, or user-approved updates.

    Expose app tools

    Use scripts/ to describe safe app endpoints and tool actions, such as browser state, account settings, open panels, logs, or retryable harness actions.

    Keep React thin

    In React apps, render slot summaries and controls from API responses. Keep permission checks, private slot reads, and agent context assembly on the server or trusted runtime.